基于小波变换和压缩感知的低速率语音编码方案  

Low bit rate speech coding based on wavelet transform and compressed sensing

作　　者：叶蕾[1] 杨震[1] 郭海燕[1] 

机构地区：[1]南京邮电大学,南京210003

出　　处：《仪器仪表学报》2010年第7期1569-1575,共7页Chinese Journal of Scientific Instrument

基　　金：国家自然科学基金(60971129);江苏省普通高校研究生科研创新计划项目(CX09B_148Z)资助

摘　　要：本文提出一种新的低速率语音编码方案,基于语音信号小波变换高频系数的稀疏性,利用压缩感知原理,将小波变换高频系数进行压缩感知投影成数据量大大减少的观测序列,然后对观测序列采用码激励线性预测技术进行编解码,根据解码后的观测序列,利用线性规划技术对小波变换高频系数进行重构,小波变换低频系数采用矢量量化技术编解码,并采用后置低通滤波器改善解码后小波高低频系数合成语音的听觉效果。该编码方案在低数码率(2.64～3.5 Kb/s)时得到的重构语音平均MOS分为3.0～3.4,达到4.8 Kb/s码激励线性预测语音编码质量。A new low bit rate speech coding technology is proposed. Based on the sparsity of high frequency wavelet transform coefficients of speech signal, compressed sensing theory is applied to project the high frequency wavelet transform coefficients to measurement sequences whose data amount is greatly decreased. The measurement sequences are coded and decoded using Codebook-Excited Linear Prediction technique. Based on the decoded measurement sequences, the reconstruction of the high frequency wavelet transform coefficients are achieved using Linear Programming; the low frequency wavelet transform coefficients are coded and decoded using Vector Quantization technique. The subject quality of the speech signal synthesized from decoded wavelet transform coefficients is improved by a low pass post-filter. The average MOS scores of the reconstruction signal are between 3.0 - 3.4 in low bit rate （2.64 - 3.5 Kb/s）, which achieves the quality of 4.8 Kb/s Codebook-Excited Linear Prediction speech coding technique.

关 键 词：小波变换 压缩感知 码激励线性预测 矢量量化 线性规划 

分 类 号：TN912.3[电子电信—通信与信息系统]